Relative Pose Estimation Pipeline for Docking of SmallSats

This work proposes an accurate real-time 6 DoF pose estimation pipeline implemented around a rendezvous and docking ground simulator for SmallSats using a stereo-camera setup. The pipeline processes synchronized stereo frames with an interface segmentation model to isolate the docking target, and masked stereo outputs are generated frame-wise for efficient downstream processing. Additionally it considers recent breakthroughs in deep learning by utilizing VSS models, specifically a VMamba-based keypoint backbone [1], in order to efficiently extract image keypoints in real-time, followed by different post-processing and robustness strategies to ensure accurate results. The prediction model is trained on a synthetic dataset generated from a Gaussian Splatting render of the docking interface, followed by dedicated preprocessing and augmentation steps. Lastly we attempt to robustly explore the correspondence hypothesis space by using homography scoring and temporal consistency, eliminating erroneous and ambiguous solutions enabling us to continuously predict the relative pose between the observed docking interface and the rest pose. The docking interface under investigation is the intelligent Space System Interface (iSSI) from iBOSS. It is primarily designed for modular satellites and OOS. Accuracy requirements are strict to mate the active iSSI module mounted on a chaser to the passive iSSI module mounted on a target. The goal of the 6DoF pose estimation pipeline is to reach a rotational precision of 0.5° and a translational precision of 1 mm at 1-2 Hz.